基于高光谱技术融合图像信息的脱绒棉种品种分类检测研究

开展种子品种的识别研究是保证种子质量的重要手段。利用高光谱图像技术融合图像特征信息对脱绒棉种的品种进行判别分析。采集4个品种共240粒脱绒棉种样本的高光谱图像数据(400～1 000 nm)，提取样本的光谱信息及长、宽、面积、圆形度、等12个形态特征。采用连续投影算法(SPA)选出11个特征波段作为输入结合偏最小二乘判别分析法(PLS-DA)、软独立模式识别法(SIMCA)、最邻近节点算法(KNN)、主成分分析结合线性判别(PCA-LDA)及二次判别(PCA-QDA)进行建模分析，得出PLS-DA建模集和预测集的总体识别率分别为93%和90%。利用图像信息进行建模分析，模型整体的识别率均不高，说明单独使用高光谱图像的形态特征进行分类效果不佳。将特征波段的光谱和形态特征信息进行融合作为输入，建立基于PLS-DA，SIMCA，KNN，PCA-LDA及PCA-QDA的信息融合模型，其精度均比基于光谱或形态信息模型高，其中PLS-DA模型识别效果最好，建模集和预测集总体识别率分别为98%和97%。表明融合高光谱图像的光谱与图像信息可以在少量波段情况下有效的提高脱绒棉种品种的分类检测精度。

The Classification of Delinted Cottonseeds Varieties by Fusing Image Information Based on Hyperspectral Image Technology

Study on identification of seed varieties is an important means of ensuring seed quality. The paper uses hyperspectral image technology and fusing image feature to identify different varieties of delinted cottonseeds. Hyperspectral image data (400～1000nm) of 4 types a total of 240 delinted cottonseeds samples were acquired. In addition, the spectral information and 12 morphological characteristics such as length width area，and circularity were extracted. Moreover, 11 effective wave-lengths(EWs) were to be selected by successive projection algorithm(SPA). And then 11 EWs of the calibration set were used as input to build a partial least quares discriminant analysis(PLS-DA)，soft independent modeling of class analogy(SIMCA)，K-nearest neighbor algorithm(KNN)，principal component analysis was combined with linear discriminant analysis (PCA-LDA) and quadratic discriminant analysis (PCA-QDA) were used to build models. The results showed that the total identification rate of the PLS-DA model were 93% for the calibration set and 90% for the prediction set, respectively. When using image information modeling analysis，the overall recognition rate of the model is not high，which showed that the effect of classification is not good when only using morphological characteristics of hyperspectral images. Then，we fused the spectral and morphological information of the feature band as input，and established the data fusion model based on the analysis of PLS-DA，SIMCA，KNN，PCA-LDA and PCA-QDA. It suggested that the data fusion model showed better performance than the individual image model and spectral model，PLS-DA model had the best recognition effect，the overall recognition rate of calibration set and prediction set was 98% and 97% respectively. The experimental results indicated that fusing the spectral and image information of hyperspectral images could effectively improved discrimination accuracy for delinted cottonseeds at the case of a small amount of wavebands.